PULSE
FEED
vulnKEV agrega CVE-2025-39964 — Linux / KernelvulnKEV agrega CVE-2026-53266 — Linux / KernelvulnKEV agrega CVE-2025-39682 — Linux / KernelvulnKEV agrega CVE-2026-58704 — Google / PixelvulnKEV agrega CVE-2026-76460 — Cisco / Identity Services EnginevulnKEV agrega CVE-2026-87886 — Acronis / BackupvulnKEV agrega CVE-2026-76461 — Cisco / Secure Email GatewayvulnKEV agrega CVE-2026-84869 — ConnectWise / ScreenConnectvulnKEV agrega CVE-2026-42016 — JFrog / ArtifactoryvulnKEV agrega CVE-2026-42018 — JFrog / ArtifactoryvulnKEV agrega CVE-2026-85706 — GitLab / Community Edition and Enterprise EditionvulnKEV agrega CVE-2026-86060 — MikroTik / RouterOSvulnKEV agrega CVE-2026-67277 — MikroTik / RouterOSvulnKEV agrega CVE-2026-19490 — Citrix / NetScalervulnKEV agrega CVE-2025-39964 — Linux / KernelvulnKEV agrega CVE-2026-53266 — Linux / KernelvulnKEV agrega CVE-2025-39682 — Linux / KernelvulnKEV agrega CVE-2026-58704 — Google / PixelvulnKEV agrega CVE-2026-76460 — Cisco / Identity Services EnginevulnKEV agrega CVE-2026-87886 — Acronis / BackupvulnKEV agrega CVE-2026-76461 — Cisco / Secure Email GatewayvulnKEV agrega CVE-2026-84869 — ConnectWise / ScreenConnectvulnKEV agrega CVE-2026-42016 — JFrog / ArtifactoryvulnKEV agrega CVE-2026-42018 — JFrog / ArtifactoryvulnKEV agrega CVE-2026-85706 — GitLab / Community Edition and Enterprise EditionvulnKEV agrega CVE-2026-86060 — MikroTik / RouterOSvulnKEV agrega CVE-2026-67277 — MikroTik / RouterOSvulnKEV agrega CVE-2026-19490 — Citrix / NetScaler
← All CVEs
CVE WatchSep 18, 2026

CVE-2026-93841

vLLM through 0.29.0 contains a memory corruption vulnerability in the Triton _bincount_kernel where prompt token IDs index the penalty promp

CVSS

3.7

Low

EPSS

KEV

Exploit Today

0-100

Published: Sep 18, 2026 · Last modified: Sep 18, 2026 · CWE-129

EPSS · 30d

Not enough EPSS history yet.

Technical description

vLLM through 0.29.0 contains a memory corruption vulnerability in the Triton _bincount_kernel where prompt token IDs index the penalty prompt-presence bitset without bounds checking against vocabulary size. Attackers can submit multimodal audio requests with tokens equal to vocabulary size, causing out-of-bounds writes that corrupt concurrent requests' sampler state and alter repetition penalty behavior.

Official references
Related CVEs
CVECVSSEPSSKEVRExploitTitleMod.
CVE-2026-938403.7 LOW
vLLM before 0.29.0 validates allowed_token_ids against tokenizer length instead of model output logits width in SamplingParams._validate_allowed_token_ids(). Attackers can supply token IDs above the output vocabulary that pass validation, causing LogitBiasState to corrupt GPU logits state and allow concurrent requests to sample tokens outside their allowlists.9h
CVE-2026-636355.5 MED
OpenImageIO is a toolset for reading, writing, and manipulating image files of any image file format relevant to VFX / animation. Prior to 3.0.21.0, 3.1.16.0, and 3.2.0.3-beta1, A crafted psd with an invalid color_mode bypasses normal validation when oiio:rawcolor or psd:rawdata is enabled. psdinput::setup() then uses the attacker-controlled value to index fixed color-mode tables, causing a global out-of-bounds read and potentially a bogus allocation, resulting in denial of service. The affected implementation is identified by src/psd.imageio/psdinput.cpp, PSDInput::validate_header(), PSDInput::setup(), oiio:RawColor, psd:RawData, color_mode, and mode_channel_count, which define the relevant source path, functions, state, and trigger. This issue is fixed in versions 3.0.21.0, 3.1.16.0, and 3.2.0.3-beta1.13h
CVE-2026-634205.5 MED
OpenImageIO is a toolset for reading, writing, and manipulating image files of any image file format relevant to VFX / animation. Prior to 3.0.21.0, 3.1.16.0, and 3.2.0.3-beta1, An indexed psd with transparency metadata creates fewer stored channel_buffers than the spec.nchannels value advertised by the rawcolor path. when oiio:rawcolor or psd:rawdata is enabled, psdinput::read_native_scanline() passes spec.nchannels to psdinput::interleave_row(), which indexes beyond channel_buffers, resulting in a heap out-of-bounds read and process crash. The affected implementation is identified by src/psd.imageio/psdinput.cpp, PSDInput::read_native_scanline(), PSDInput::interleave_row(), oiio:RawColor, psd:RawData, channel_buffers, and spec.nchannels, which define the relevant source path, functions, state, and trigger. This issue is fixed in versions 3.0.21.0, 3.1.16.0, and 3.2.0.3-beta1.13h
CVE-2026-935927.5 HIG
vLLM versions before 0.28.0 fail to validate the lower bound of token IDs in the /v1/embeddings and /pooling endpoints, allowing unauthenticated attackers to crash the engine by submitting negative token IDs. A single request with a negative token ID triggers a CUDA device-side assertion that poisons the GPU context, causing all subsequent requests to fail until the process restarts.11h
CVE-2026-91101
21.3%
6HP has identified and remediated multiple externally reported vulnerabilities within HPLIP. The findings affect several software components that could potentially enable remote code execution, privilege escalation, denial of service, information disclosure, or unauthorized file modification under certain conditions.2d
CVE-2026-552099.8 CRI
36.0%
11resdata is software for reading and writing result files from the Eclipse reservoir simulator. Prior to 6.2.9, resdata insufficiently validates numeric fields, grid dimensions, keyword sizes, and array indexes while parsing untrusted GRDECL files in lib/resdata/rd_kw_grdecl.cpp and lib/resdata/rd_grid.cpp. Malformed COORD, ZCORN, CORSNUM, ACTNUM, or MAPAXES data can reach rd_grid_alloc_GRDECL_kw__ with inconsistent lengths, while unbounded floating-point conversion can exceed the intended parser buffer. In a network service that accepts untrusted GRDECL files, these conditions can cause a classic buffer overflow, out-of-bounds reads, invalid array access, NULL pointer dereference, memory corruption, or service termination. This issue is fixed in version 6.2.9.4d